Re: GGC ICEs

[Mike Stump <mrs at apple dot com>]

On Jan 16, 2007, at 1:50 PM, Jan Hubicka wrote:
> Also people complain that ggc_free increase memory fragmentation and
> slows compiler down (I didn't see any support from benchmark for this
> however).

Hum...  I do make -j6 builds on a 4 CPU, 1 GB machine that starts off  
wanting to use 1G for the OS, the filesystem cache for the gcc source  
tree and the local file system cache for svn.  Ok, stop laughing at  
me now.  If I have to page in data that had been evicted, the cost is  
about 10ms per continuous group paged back in.  We have to be  
somewhat nice to memory so that the total system performance doesn't  
suffer too much.  At the end of the day, modern boxes are  
occasionally used for other things besides compiling.  I'd like to  
think that a single page in can cover all the cost of all calls to  
ggc_free.  I'd love to see people settle this by posting numbers, if  
there is doubt.  Certainly I'm more interested in fast turn around  
during compilation than anything else.  I'd like to think that not  
everyone else has 8GB machines already.

> It can be, but we would have to discuss the fragmentation issues and
> possibility of making production compiler more often ICEing.

Ah, I think I get what you're proposing.  A way to verify it is free  
without actually freeing it, yet.  The idea being, after enough time  
passes and it seems safe enough (or maybe just to wait for stage 1  
again), then turn it into a real ggc_free call.?  Ah, that sounds  
reasonable to me.  I think I missed the conservatism of the plan,  
sorry about that.